System and method for traffic sign recognition

• US 8,233,670 B2
• Filed: 12/31/2007
• Issued: 07/31/2012
• Est. Priority Date: 09/13/2007
• Status: Expired due to Fees

• Current Assignee: Cognex Corporation
• Original Assignee: Cognex Corporation
• Inventors: Moed, Michael C., Bonnick, Thomas W.
• Primary Examiner(s): Le, Brian Q
• Assistant Examiner(s): BROOKS, JULIAN D
• Application Number:
  US11/967,966
• Publication Number:
  US 20090074249A1
• Time in Patent Office: 1,674 Days
• Field of Search: 382/100, 382/103, 382/113, 382/209, 382/181, 382/224, 382/159, 382/226, 382/254, 382/274
• US Class Current: 382/113
• CPC Class Codes: G06V 20/582 of traffic signs

Abstract

This invention provides a vehicle-borne system and method for traffic sign recognition that provides greater accuracy and efficiency in the location and classification of various types of traffic signs by employing rotation and scale-invariant (RSI)-based geometric pattern-matching on candidate traffic signs acquired by a vehicle-mounted forward-looking camera and applying one or more discrimination processes to the recognized sign candidates from the pattern-matching process to increase or decrease the confidence of the recognition. These discrimination processes include discrimination based upon sign color versus model sign color arrangements, discrimination based upon the pose of the sign candidate versus vehicle location and/or changes in the pose between image frames, and/or discrimination of the sign candidate versus stored models of fascia characteristics. The sign candidates that pass with high confidence are classified based upon the associated model data and the drive/vehicle is informed of their presence. In an illustrative embodiment, a preprocess step converts a color image of the sign candidates into a grayscale image in which the contrast between sign colors is appropriate enhanced to assist the pattern-matching process.

Claims

1. A vehicle-mounted traffic sign recognition system comprising:

a forward-looking camera configured to capture stream data of imagery containing image frames having traffic sign images therein;

a rotation and scale-invariant geometric pattern-matching process constructed and arranged to identify traffic sign candidates in the stream data, the geometric pattern-matching process including a pre-training mechanism to pre-train with predetermined traffic sign models having exterior geometric features and to classify each of the traffic sign candidates as either an unrecognized object or as a sign of a predetermined type upon a high-confidence match between each of the traffic sign candidates and at least one of predetermined traffic sign models, and wherein the geometric pattern-matching process also produces pose information concerning the traffic sign candidates including at least one of a position, scale, skew and rotation with respect to at least one of the predetermined traffic sig models matched; and

with a plurality of discrimination processes, analyzing, in succession, each of the traffic sign candidates received from the traffic sign candidate producer with respect to pre-trained traffic sign features to either increase or decrease a confidence level in each of the traffic sign candidates, and passing or failing each of the traffic sign candidates based upon the confidence level, and associating passed traffic sign candidates with a predetermined type of traffic sign for use by the method, the plurality of discrimination processes including at least two of;

a) a color discrimination process that compares a color in a closed geometric shape of each of the traffic sign candidates to color features in an associated model from the pre-trained traffic sign features,

b) a pose discrimination process that compares a pose of each of the traffic sign candidates to pose features in an associated model from the pre-trained traffic sign features, and

c) a fascia discrimination process that compares predetermined fascia features of each of the traffic sign candidates in a current pose to fascia features in an associated model from the pre-trained traffic sign features to determine a subtype of a sign candidate based upon an identified fascia matching the model from the pre-trained traffic sign features.

2. The system as set forth in claim 1 wherein the geometric pattern-matching process includes a pre-training mechanism to pre-train with predetermined traffic sign models having interior geometric features and to classify predetermined of the traffic sign candidates as either the unrecognized object or as the sign of the predetermined type upon a high-confidence match between the predetermined of the traffic sign candidates and at least one of predetermined traffic sign models having the interior geometric features.

3. The system as set forth in claim 1 further comprising one or more discrimination processes that each analyze each of the traffic sign candidates received from the geometric pattern-matching process to either increase or decrease a confidence level in each of the traffic sign candidates, and a confidence level determination process that passes or fails the sign candidates based upon the confidence level.

4. The system as set forth in claim 3 further comprising a preprocessor that receives the image frames as color imagery and that computes grayscale intensity from colors in the image frames and that enhances contrast between predetermined color ranges and hues, and generates corresponding enhanced grayscale intensity image frames for use by the geometric pattern-matching process.

5. The system as set forth in claim 3 wherein the one or more discrimination processes include a color discrimination process that compares a color in a closed geometric shape of each of the traffic sign candidates to color features in an associated model from the pre-trained traffic sign features.

6. The system as set forth in claim 3 wherein the one or more discrimination processes include a pose discrimination process that compares a pose of each of the traffic sign candidates to pose features in an associated model from the pre-trained traffic sign features.

7. The system as set forth in claim 3 wherein the one or more discrimination processes include a fascia discrimination process that compares predetermined fascia features of each of the traffic sign candidates in a current pose to fascia features in an associated model from the pre-trained traffic sign features and that is constructed and arranged to determine a subtype of a sign candidate based upon an identified fascia matching the model from the pre-trained traffic sign features.

8. The system as set forth in claim 3 wherein the classification mechanism is operatively connected to at least one of (a) an alert system that reports the predetermined class of each of the passed traffic sign candidates to a vehicle driver, (b) a vehicle control system that employs data related to the predetermined class in controlling operation of the vehicle;

and (c) a recording system that stores predetermined data related to the sign candidates.

9. The system as set forth in claim 3 further comprising a traffic sign character and symbol recognition process that identifies characters and symbols contained within each of the candidate traffic signs and reports data associated with the identified characters and symbols.

10. The system as set forth in claim 1 further comprising a preprocessor that receives the image frames as color imagery and that computes grayscale intensity from colors in the image frames and that enhances contrast between predetermined color ranges and hues, and generates corresponding enhanced grayscale intensity image frames for use by the geometric pattern-matching process.

11. The system as set forth in claim 10 wherein the contrast is enhanced based upon predetermined intensity data associated with a confusable combination of the colors in the image frames and based upon predetermined color hues present in the image frames.

12. A method for recognizing traffic signs from a vehicle comprising the steps of:

capturing stream data of imagery containing image frames having traffic sign images therein;

identifying, with a rotation and scale-invariant geometric pattern-matching process, traffic sign candidates in the stream data, the geometric pattern-matching process including a pre-training mechanism to pre-train with predetermined traffic sign models having exterior geometric features and classifying each of the traffic sign candidates as either an unrecognized object or as a sign of a predetermined type upon a high-confidence match between the each of the traffic sign candidates and at least one of predetermined traffic sign models, and wherein the geometric pattern-matching process also produces pose information concerning the traffic sign candidates including at least one of a position, scale, skew and rotation with respect to at least one of the predetermined traffic sign models matched; and

with a plurality of discrimination processes, analyzing, in succession, each of the traffic sign candidates received from the traffic sign candidate producer with respect to pre-trained traffic sign features to either increase or decrease a confidence level in each of the traffic sign candidates, and passing or failing each of the traffic sign candidates based upon the confidence level, and associating passed traffic sign candidates with a predetermined type of traffic sign for use by the method, the plurality of discrimination processes including at least two of;

a) a color discrimination process that compares a color in a closed geometric shape of each of the traffic sign candidates to color features in an associated model from the pre-trained traffic sign features,

b) a pose discrimination process that compares a pose of each of the traffic sign candidates to pose features in an associated model from the pre-trained traffic sign features, and

c) a fascia discrimination process that compares predetermined fascia features of each of the traffic sign candidates in a current pose to fascia features in an associated model from the pre-trained traffic sign features to determine a subtype of a sign candidate based upon an identified fascia matching the model from the pre-trained traffic sign features.

13. The method as set forth in claim 12 wherein the geometric pattern-matching process includes a pre-training mechanism to pre-train with predetermined traffic sign models having interior geometric features and to classify predetermined of the traffic sign candidates as either the unrecognized object or as the sign of the predetermined type upon a high-confidence match between the predetermined of the traffic sign candidates and at least one of predetermined traffic sign models having the interior geometric features.

14. The method as set forth in claim 13 further comprising, with one or more discrimination processes, analyzing each of the traffic sign candidates received from the geometric pattern-matching process with respect to pre-trained traffic sign features to either increase or decrease a confidence level in each of the traffic sign candidates, and passing or failing each of the traffic sign candidates based upon the confidence level, and associating passed traffic sign candidates with a predetermined type of traffic sign for use by the system.

15. The method as set forth in claim 14 further comprising preprocessing the image frames received as color imagery including computing grayscale intensity from colors in the image frames and enhancing contrast between predetermined color ranges and hues, and generating corresponding enhanced grayscale intensity image frames for use by the geometric pattern-matching process.

16. The method as set forth in claim 14 wherein the one or more discrimination processes include a color discrimination process that compares a color in a closed geometric shape of each of the traffic sign candidates to color features in an associated model from the pre-trained traffic sign features.

17. The method as set forth in claim 14 wherein the one or more discrimination processes include a pose discrimination process that compares a pose of each of the traffic sign candidates to pose features in an associated model from the pre-trained traffic sign features.

18. The method as set forth in claim 14 wherein the one or more discrimination processes include a fascia discrimination process that compares predetermined fascia features of each of the traffic sign candidates in a current pose to fascia features in an associated model from the pre-trained traffic sign features to determine a subtype of a sign candidate based upon an identified fascia matching the model from the pre-trained traffic sign features.

19. The method as set forth in claim 14 further comprising at least one of (a) reporting the predetermined class of each of the passed traffic sign candidates to a vehicle driver, and (b) employing data related to the predetermined class in controlling operation of the vehicle and (c) storing predetermined data related to the sign candidates.

20. The method as set forth in claim 14 further comprising identifying, with a traffic sign character and symbol recognition process, characters and symbols contained within each of the candidate traffic signs and reports data associated with the identified characters and symbols.

21. The method as set forth in claim 12 further comprising preprocessing the image frames received as color imagery including computing grayscale intensity from colors in the image frames and enhancing contrast between predetermined color ranges and hues, and generating corresponding enhanced grayscale intensity image frames for use by the geometric pattern-matching process.

22. The method as set forth in claim 21 wherein the step of enhancing is based upon predetermined intensity data associated with a confusable combination of the colors in the image frames and based upon predetermined color hues present in the image frames.